modelling of air temperature using ann and remote sensing

14
This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright

Upload: mehmet-sahin

Post on 22-Jan-2018

156 views

Category:

Engineering


2 download

TRANSCRIPT

Page 1: Modelling of air temperature using ann and remote sensing

This article appeared in a journal published by Elsevier. The attachedcopy is furnished to the author for internal non-commercial researchand education use, including for instruction at the authors institution

and sharing with colleagues.

Other uses, including reproduction and distribution, or selling orlicensing copies, or posting to personal, institutional or third party

websites are prohibited.

In most cases authors are permitted to post their version of thearticle (e.g. in Word or Tex form) to their personal website orinstitutional repository. Authors requiring further information

regarding Elsevier’s archiving and manuscript policies areencouraged to visit:

http://www.elsevier.com/copyright

Page 2: Modelling of air temperature using ann and remote sensing

Author's personal copy

Modelling of air temperature using remote sensing and artificialneural network in Turkey

Mehmet S�ahin ⇑

Engineering Faculty, Siirt University, 56100, Siirt, Turkey

Received 17 March 2012; received in revised form 14 June 2012; accepted 16 June 2012Available online 26 June 2012

Abstract

The aim of this research was to forecast monthly mean air temperature based on remote sensing and artificial neural network (ANN)data by using twenty cities over Turkey. ANN contained an input layer, hidden layer and an output layer. While city, month, altitude,latitude, longitude, monthly mean land surface temperatures were chosen as inputs, and monthly mean air temperature was chosen asoutput for network. Levenberg–Marquardt (LM) learning algorithms and tansig, logsig and linear transfer functions were used in thenetwork. The data of Turkish State Meteorological Service (TSMS) and Technological Research Council of Turkey–Bilten for the periodfrom 1995 to 2004 were chosen as training when the data of 2005 year were being used as test. Result of research was evaluated accordingto statistical rules. The best linear correlation coefficient (R), and root mean squared error (RMSE) between the estimated and measuredvalues for monthly mean air temperature with ANN and remote sensing method were found to be 0.991–1.254 K, respectively.� 2012 COSPAR. Published by Elsevier Ltd. All rights reserved.

Keywords: Air temperature; Artificial neural network; NOAA; AVHRR; Remote sensing; Satellite

1. Introduction

Air temperature is a measure of how hot or cold the airis. It is the most commonly measured by weather parame-ter. More specifically, air temperature describes the kineticenergy or energy of motion of the gases that make up air. Ifgas molecules move more quickly, air temperatureincreases otherwise it decreases. The estimation of air tem-perature is useful for lots of applications including study ofvector-borne diseases (Thomson et al., 1996; Goetz et al.,2000), epidemic forecasting (Bian et al., 2006), weatherforecasting, veterinary uses, climate change (Kuchariket al., 2010; Bocchiola and Diolaiuti, 2010; Kittel et al.,2011), determination of various heat and radiation fluxes(Brunsell et al., 2011), vapour pressure deficit, water poten-tial (Aasamaa and Sober, 2011), urban land use and urban

heat island (Cheval et al., 2009), shortwave and longwaveradiation (Stanelle et al., 2010), stomatal resistance (Leeet al., 2011), ecology (Myint et al., 2010; Smith, 2011; Hed-ing et al., 2011), hydrology (Jain et al., 2011) and atmo-spheric sciences. And also knowledge of air temperatureis necessary for the health of human being (Elwood et al.,1993; Kunst et al., 1993; Ballester et al., 1997; Analitiset al., 2008; Michelozzi et al., 2009; Almeida et al., 2010)

Generally, air temperature which is very important is thefrequently observed and recorded by weather meteorologi-cal stations with high accuracy. But density of the stationnetwork is normally not sufficient when air temperatureis employed in regional numerical models for climate orevapotranspiration. So, a new method is necessary to getair temperature over the wide fields by using satellites data.The data of Advanced Very High Resolution Radiometer(AVHRR), Geostationary Operational EnvironmentalSatellite (GOES), Meteosat, TIROS Operational VerticalSounder (TOVS), LANDSAT/TM and Moderate Resolu-tion Imaging Spectroradiometer (MODIS) have been used

0273-1177/$36.00 � 2012 COSPAR. Published by Elsevier Ltd. All rights reserved.

http://dx.doi.org/10.1016/j.asr.2012.06.021

⇑ Tel.: +90 (484) 223 12 24; fax: +90 (484) 223 66 31.E-mail addresses: [email protected], sahanmehmet2000@yahoo.

com.

www.elsevier.com/locate/asr

Available online at www.sciencedirect.com

Advances in Space Research 50 (2012) 973–985

Page 3: Modelling of air temperature using ann and remote sensing

Author's personal copy

in historical term. The air temperatures have beenestimated from land surface temperatures retrieved fromsatellite images with 1 km � 1 km or 4 km � 4 km resolu-tion by using thermal infrared radiation emitted from thesurface by researchers. Generally, land surface temperaturehas been retrieved from two thermal infrared bands (chan-nels 4 and 5 of AVHRR of NOAA) located at 11 lm and12 lm by using split-window equations. The split-windowalgorithms are belonged to the difference in the brightnesstemperatures of thermal infrared bands. Furthermore, landsurface temperature depends on the magnitude of the dif-ference between the two grounds emissivity in the bands(Becker, 1987).

In the literature, Kawamura and Edamatsu (1993) usedAVHRR thermal infrared data and estimated the air-tem-perature (Ta) with an rms error 2.5–2.8 K. Prihodko andGoward (1997) explored a methodology for estimatingair temperature directly from remotely sensed observationsusing the correlation between a spectral vegetation indexand surface temperature (temperature-vegetation index).These air temperature estimates were compared with coin-cident ground-measured air temperatures recorded at stan-dard meteorological stations. A strong correlation(R = 0.93) was found between the satellite estimates andmeasured air temperatures with a mean error of 2.92 K.The correlation between surface temperature and spectralvegetation index has been obtained to estimate air temper-ature from satellite images that was called the Surface Tem-perature/Spectral Vegetation Index (TVX) by Czajkowskiet al. (1997), Prihodko and Goward (1997). The TVX con-cept and the empirical method have provided estimated airtemperatures with an rms error of 3 K (Cresswell et al.,1999). Lakshmi et al. (2001) evaluated the ability of satel-lites to map air temperature over the large land surfaceareas. Then, they developed an algorithm that derives sur-face air temperature by using observations from the TOVSsuite of instruments and also from the AVHRR. The resultof their study showed that the average bias over the 3-months period compared with ground-based observationswas approximately 2 K or less for the three times of daywith TOVS having lower biases than AVHRR. Ridderingand Queen (2006) presented a technique for producing esti-mates of near surface air temperature in complex terrainbased on composite data from the NOAA-AVHRR.Results were tested against meteorological data. The corre-lation coefficient approximately 0.742 and standard errorsof the estimate of 2.73 K were observed in the final modelimplementation. Stisen et al.(2007) obtained air tempera-ture belonging to a model that was including satellite datawith high temporal resolution that were desired for severalmodelling applications by exploiting the thermal split-win-dow channels in combination with the red and near infra-red channels of the geostationary MSG SEVIRI sensor.The research results showed that accuracy of estimationof the air temperature was changing from 2.55 K to2.99 K by root mean squared error. Vancutsem et al.(2010) studied to explore the possibility of retrieving

high-resolution near surface air temperature (Ta) data fromthe MODIS connected to land surface temperature(LST) over different ecosystems in Africa. The comparisonsbetween night MODIS-LST data with minimum Ta

showed that MODIS night-time products haveprovided a good estimation of minimum Ta (standarddeviation = 2.4 K).

The other way of estimation air temperature is ANNwhich has been used for various purposes in remote sens-ing. Jang et al. (2004) employed multilayer feed-forwardneural networks to estimate air temperatures in SouthernQuebec (Canada) using AVHRR images. The input vari-ables for the networks were the five bands of the AVHRRimage, surface altitude, solar zenith angle, and Julian day.The estimation was carried out using a dataset collectedduring the growing season from June to September 2000.Levenberg–Marquardt back-propagation (LM-BP) wasused to train the networks. The early stopping methodwas applied to improve the LM-BP and to generalize thenetworks. The network using all five bands, Julian day,altitude, and solar zenith angle provided the best results,with 22 nodes in the hidden layer. The difference betweenestimated and station air temperatures was obtained within1.79 K by RMSE.

In this study, ANN was applied to estimate air temper-atures of Adana, Afyonkarahisar, Ankara, Antalya, Art-vin, Balıkesir, Denizli, Erzurum, Eskis�ehir, _Istanbul-Goztepe, _Izmir, Kars, Kayseri, Konya, Malatya, Rize,Samsun, Sivas, S�anlıurfa and Van cities. For this purpose,raw data of NOAA-AVHRR were converted into Level 1Bdata set by using “Quorum to Level1B” software. Then alldata sets were corrected radiometrically, geometrically andatmospherically and used these data sets to get brightnesstemperatures of band 4 and band 5 of NOAA-AVHRRimages. In continuance, brightness temperatures wereemployed to estimate land surface temperature. The city,month, altitude, latitude, longitude, monthly mean landsurface temperature were chosen as input data whilemonthly mean air temperature was chosen as output datain artificial neural network. The LM learning algorithmsand tansig, logsig and linear transfer functions were usedin the network. Results of the network have indicated thatmonthly mean air temperature can be estimated correctlyin Turkey by using ANN which is including meteorologicaland satellite data. At the same time, the study showed thatthe model M8/6-14-1 which developed over Turkey hasprovided more accurate outcomes than other studies inthe literature.

2. Study areas and their characteristics, data sources

In the study, cities of Adana, Afyonkarahisar, Ankara,Antalya, Artvin, Balıkesir, Denizli, Erzurum, Eskis�ehir,_Istanbul-Goztepe, _Izmir, Kars, Kayseri, Konya, Malatya,Rize, Samsun, Sivas, S�anlıurfa and Van were chosen asstudy areas (see Fig. 1 and see Table 1). The mentioned cit-ies have various climatic conditions from one another. The

974 M. S�ahin / Advances in Space Research 50 (2012) 973–985

Page 4: Modelling of air temperature using ann and remote sensing

Author's personal copy

climate of Adana, Antalya carries the properties ofMediterranean climate; too hot and arid in summers whilewarm and rainy in winters. The highlands have a mixedproperty of Mediterranean climate and continental climate.Precipitation is generally in the form of rainfall. The 29%of the city lands is covered with forests. The forests areon the highlands. The flora is composed of Mediterraneanplants; with scrubs up to 700–800 m altitude on mountainslopes while with larches and cedars on the uplands.

Ankara, Eskis�ehir, Kayseri, Konya, Sivas have terres-trial climate that is the most pluvial season in spring. Theclimate conditions and topographic structure have enabledthe growth of two plant associations around cities; steppeand forest. The most prevalent plant association in theregions is steppe. The steppe is common on plateaus andin valleys where drop little precipitation. There is almostno tree in that plant association. For the most part, thornybushes are seen. In addition, angustifolia, willow, and

populous trees in rows all along the stream are in steppeclimate.

Since Balıkesir and _Istanbul-Goztepe are under theimpact of Marmara, Mediterranean and terrestrial climate,a plant association at a quarter of the cities is not seen inother part of the region. The 30% of the surface is sylvanwhereas 32% of the city lands is pasture area, and more,23% arable lands; 15% olive groves, orchards, and vegeta-ble gardens.

Mediterranean climate is dominant in Denizli and _Izmirwhereas Afyonkarahisar has terrestrial climate; hot andarid in summers while rainy and warm in winters. The hot-test months are July and August while the coldest monthsare January and February. There is almost no snowfall.Nearly 50% of the city land is scrub and sylvan while33% planted area, and 15% pasture area.

The climate of Artvin, Rize and Samsun has variationsat coastal lines and inlands. Though there is a dominantMediterranean climate at coastal lines, due to mountains,the inlands are under the influence of terrestrial climate.Although it is not too hot in summers, the rate of humidityis rather high on account of dam reservoirs as well as sea-water evaporation. The winter months are slightly cold andfairly rainy. The precipitation is mostly rainfall at coastalareas while the inlands experience more snowy days, andunder the effect of continental climate. The land surfaceis encased in plains, orchards, gardens, pasture areas andplanted fields. The mountains are brimming with forests.

Terrestrial climate dominates in Malatya and S�anlıurfa.Summers are long lasting and hot while winters are verycold. Temperature difference between night and day ishigh. The 60% of the city lands is planted while 38% is pas-ture area. Field of forest and scrub is rare; 0.6%. The citylands are in the form of steppe.

Erzurum is one of the highest and coldest cities of Tur-key. Harsh continental climate rules over the city. The win-ters are very cold whereas very hot and arid in summers.Visual appearance of the city is green in springs, white inwinters, yellow (steppe) in falls and summers. The areawith forest and scrub covers 9%. Scotch pine and oak

Fig. 1. Air temperature measuring stations in Turkey.

Table 1Geographical parameters for the stations.

Stations Latitude (�N) Longitude (�E) Altitude (m)

Adana 37.03 35.21 27Afyonkarahisar 38.44 30.33 1035Ankara 39.57 32.53 891Antalya 36.42 30.44 64Artvin 41.11 41.49 628Balıkesir 40.06 27.39 37Denizli 37.47 29.05 425.29Erzurum 39.57 41.4 1758.18Eskis�ehir 39.45 30.33 805_Istanbul-Goztepe 40.58 29.05 32.98_Izmir 38.23 27.04 28.55Kars 40.37 43.06 1775Kayseri 38.43 35.29 1092Konya 37.52 32.28 1030Malatya 38.21 38.13 947.87Rize 41.02 40.30 8Samsun 41.21 36.15 4Sivas 39.45 37.01 1285S�anlıurfa 37.09 38.47 547.18Van 38.28 43.21 1670.58

M. S�ahin / Advances in Space Research 50 (2012) 973–985 975

Page 5: Modelling of air temperature using ann and remote sensing

Author's personal copy

consist of the area at 1900–2000 m altitude. Pasture areatakes place 68% of the city land while 18% planted area.

Kars and Van are another cities dominating terrestrialclimate. The winters are harsh and long lasting. The LakeVan provides less harsh winter conditions on very highzones of Van. The summers are slightly rainy and veryhot. Even if myriad plantation associations are seen aroundVan and Kars; plains, the mountains are generally treeless.The 70% of the city lands is pasture area while 23% isplanted and 2% is covered with forest and scrub (http://www.cografya.gen.tr/).

As known, air temperature is measured 2 m above fromthe ground as World Meteorological Organization hasdetermined and made it a guideline. In the study; altitude,latitude, longitude and air temperature data were takenfrom TSMS for the period from 1995 to 2005 while satellitedata were obtained from Scientific and TechnologicalResearch Council of Turkey-Bilten, simultaneously.

3. Methodology

3.1. Split-window method

To retrieve LST from the AVHRR top-of-atmospherebrightness temperature measurements, it was chosen assplit window algorithm. This approach is mainly basedon adjacent thermal channels of satellite and an empiricalinverse model, and it is derived from a first order expansionof the Taylor series applied to the Planck function. In thecase of the AVHRR sensors, channels 4 and 5 are used.The algorithm assumes that atmospheric attenuation (duemostly to atmospheric water vapour) is greater in channel5 than in channel 4, and the difference in measured radi-ance between the two channels increases with increasingwater vapour (Pinheiro et al., 2006).

Firstly raw data of NOAA12-14-15/AVHRR which hadno cloud, were translated into Level-1B format by usingQuorum Software and in second step, brightness tempera-ture of channel-4 and channel-5 (range 10.3–11.3 lm andrange11.5–12.5 lm, respectively) were obtained fromLevel-1B data by making use of Envi 4.3 image-processingprogrammer. Then, radiometric and geometric calibrationswere applied to the images to correct the deficiencies andflaws in the imaging sensors of the satellite. The imageswere then turned into brightness temperatures which arenecessary for the retrieval of LST from AVHRR data.

In this study, split window LST algorithm developed byUlivieri et al. (1994) was used due to its simplicity, robust-ness and superior performance in independent tests (Beckerand Li, 1995; Vazquez et al., 1997; Yu et al., 2008).

Ulivieri’s algorithm can be written as,

T ¼ T 4 þ 1:8ðT 4 � T 5Þ þ 48ð1� eÞ � 75De ð1Þ

e ¼ e4 þ e5

2ð2Þ

De ¼ e4 � e5 ð3Þ

where T4 and T5 are the brightness temperatures ofAVHRR channels 4 and 5 respectively, and, Eq. (1) wasdeveloped for cases of column atmospheric water vapourless than 3.0 g/cm2, a reasonable condition for much ofthe semi-arid portions of continental Africa (Pinheiroet al. 2006). e and De are mean spectral emission coefficientand difference of the emission coefficients for channels 4and 5, respectively.e4 and e5 are the surface emission coef-ficients which were estimated from atmospherically cor-rected NDVI (Normalized Difference Vegetation Index),using the equations given by Valor and Casselles (1995)for channel 4 and channel 5, respectively .

e4 ¼ 0:9897þ 0:029 lnðNDVIÞ ð4Þe4 � e5 ¼ 0:01019þ 0:01344 lnðNDVIÞ ð5Þwhere NDVI is a simple graphical indicator that can beused to analyse remote sensing measurements and assesswhether the target being observed contains live green veg-etation or not. Live green plants absorb solar radiationin the photosynthetically active radiation spectral region,which they use as a source of energy in the process of pho-tosynthesis. Leaf cells have also evolved to scatter (i.e., re-flect and transmit) solar radiation in the near-infraredspectral region (which carries approximately half of the to-tal incoming solar energy) because the energy level per pho-ton in that domain (wavelengths longer than about700 nm) is not sufficient to be useful to synthesize organicmolecules. A strong absorption at these wavelengths wouldonly result in over-heating the plant and possibly damagingthe tissues. Hence, live green plants appear relatively darkin the photosynthetically active radiation and relativelybright in the near-infrared (Gates, 1980). By contrast,clouds and snow tend to be rather bright in the red (as wellas other visible wavelengths) and quite dark in the near-infrared. The pigment in plant leaves, chlorophyll, stronglyabsorbs visible light (from 0.4 to 0.7 lm) for use in photo-synthesis. The cell structure of the leaves, on the otherhand, strongly reflects near-infrared light (from 0.7 to1.1 lm). The more leaves a plant has, the more these wave-lengths of light are affected. Since early instruments ofearth observation, such as NOAA’s AVHRR, acquireddata in visible and near-infrared, it was natural to exploitthe strong differences in plant reflectance to determine theirspatial distribution in this satellite images. The NDVI iscalculated from these individual measurements as follows:

NDVI ¼ NIR� VISNIRþ VIS

ð6Þ

where VIS and NIR stand for the spectral reflectance mea-surements acquired in the visible (red) and near-infrared re-gions, respectively (Goward et al., 1991; Santos and Negri,1997).

3.2. Artificial neural network

ANN is a branch of artificial intelligence which wasdeveloped in 1950s in order to imitate the biological

976 M. S�ahin / Advances in Space Research 50 (2012) 973–985

Page 6: Modelling of air temperature using ann and remote sensing

Author's personal copy

structure of the human brain (Viotti et al., 2002). TheANN models work like a black box without requiring thedetailed information of a system. Instead of requiring thisinformation, they learn the relation between the inputparameters and the controlled and uncontrolled variablesby studying the previously recorded data like non-linearregression. One more advantage of using ANN is the capa-bility of managing large and complex systems with a vastnumber of interrelated parameters. On the other hand theydon’t take into account the excess data which is veryimportant (Kalogirou, 2001).

The use of the ANNs for modelling and prediction pur-poses is increasingly becoming popular in the last decades(Cam et al., 2005). Researchers have been applying theANN method successfully in various fields of mathematics,engineering, medicine, economics, meteorology, psychol-ogy, neurology, as well as in the prediction of mineralexploration sites, in electrical and thermal load predictions,in adaptive and robotic control, and many other subjects.ANNs have been trained to overcome the limitations ofthe conventional approaches to solve complex problems.This method learns from given examples by constructingan input-output mapping in order to perform predictions(Kalogirou, 2000). Fundamentals processing element of aneural network is a neuron. Each neuron computes aweighted sum of its input signals.

The neuronal model of Fig. 2 includes an externallyapplied bias, denoted by bk. The bias bk has the effect ofincreasing or lowering the net input of the activation func-tions, depending on whether it is positive or negative,respectively.

In mathematical terms, a neuron k may be described bywriting the following pair of equations:

uk ¼Xm

j¼1

wkjxj ð7Þ

yk ¼ uðuk þ bkÞ ð8Þ

where x1,x2, . . .,xm are the inputs signals; wk1, wk2, . . ., wkm

are the synaptic weights of neuron k; uk is the linear com-biner output due to the input signals; bk is the bias; /(�) isthe activation function and yk is the output signal of theneuron. The use of bias bk has the effect of applying an

affine transformation to the output uk of the linear com-biner in the model of Fig. 2, as shown by

vk ¼ uk þ bk ð9Þ

In particular, depending on whether the bias bk is posi-tive or negative, the relationship between the induced localfield or activation potential vk of neuron k and the linearcombiner output uk is modified in the manner illustratedin Fig. 3; hereafter the term “induced local field” is used.Note that as a result of this affine transformation, thegraph of vk versus uk no longer passes through the origin.

The bias bk is an external parameter of artificial neuronk. It may be accounted for its presence as in Eq. (8). Equiv-alently, the combination of Eqs. (7) and (8) may be formu-lated as follows, respectively (Haykin, 1999):

vk ¼Xm

j¼0

wkjxj ð10Þ

yk ¼ uðvkÞ ð11Þ

The tangent sigmoid transfer function Eq. (12), log-sig-moid transfer function Eq. (13) and linear transfer functionEq. (14) are described with the following equation respec-tively (Vogl et al.,1988).

uðxÞ ¼ 2

1þ e�2xð12Þ

uðxÞ ¼ 1

1þ e�xð13Þ

uðxÞ ¼ linearðxÞ ¼ x ð14Þ

An ANN is organized as layers of neurons. Each neuronin a layer is connected to all neurons in the previous layer.An example of this type of arrangement, which is used alsoin this study, is shown in Fig. 4. The network consists of aninput layer, one hidden layer and an output layer.

When Fig. 4 examined, it can be seen that the parame-ters of city, month, altitude, latitude, longitude, andmonthly mean land surface temperatures in the input layer

x1

x2

xn

wk1

wk2

wkn

Synaptic weights

Inputs

yk

Activationfunction

Summingjunction

Output (.)ϕ

Biasbk

vk

.

.

.

.

.

.

Fig. 2. Nonlinear model of a neuron.

Linear conbiner’s output, uk

Induced local field, vk

0

Bias bk>0

bk=0

bk<0

Fig. 3. Affine transformation produced by the presence of a bias; note thatvk = bk at uk = 0.

M. S�ahin / Advances in Space Research 50 (2012) 973–985 977

Page 7: Modelling of air temperature using ann and remote sensing

Author's personal copy

were employed to able to get monthly air temperature inthe output layer. Even if the values of land surfacetemperature and air temperatures differ from one another,both are always in the case of thermal interaction. The landsurface that is heated by solar radiation warms up the air;heat lost to the air causes land surface temperature dimin-ishing. So, ANN is an indispensable parameter in estima-tion of air temperature. Further, because atmosphere isheated by the radiations reflected from the earth ground,lower atmosphere is hot while upper atmosphere is cold.There is a progressive cooling of 1 �C for every 200 m rosein the atmosphere. So, the altitude where the point of airtemperature estimation exists is crucial.

The angle at which solar radiation strikes the land sur-face is the most important factor that affects temperaturepattern on the earth. The more the angle of solar radiationincreases the direction at a right angle, the more the pointwhere it strikes gets warmer. As the strike angle contracts,the heating lessens. As known, the angle of incoming solarradiation changes with latitude. The sun shines straightdown like near the equator whereas inclined around thepoles. The temperature diminishes when moved from theequator towards the poles. That law put forwards that lat-itude is salient in calculating air temperature. Moreover,the hitting angle of solar radiation shifts throughout theyear depending on axial tilt and earth annual move. Thatincreases the importance of temporal resolution in estima-tion of air temperature accurately. In the study, monthlytemporal resolution followed and month was used as oneof the input parameters.

The longitude is an angular measurement of any pointfrom starting meridian. The only impact of longitude isthe building local time differences. As known, each pixelin the satellite image represents for a different point. Tem-perature values scattered from thousands of differentpoints reach to satellite sensors depending on differentangles. The angles were linked with longitude angles, solongitude was employed as input parameter in the study.

The other input parameter in the study is city. In sunnyand hot days, cities, in which tall buildings exist and whichare densely populated, is hotter when compared to theirsurroundings because of urban heat island (UHI) effect.

UHI effects come into existence in cities, and the tempera-ture values in these kinds of cities are higher because theyspread more heating energy. UHI effect changes from cityto city. In the study UHI effect is stated as city.

3.3. Evaluation of the estimation results

The choice of the relevant criteria allowing performanceevaluation of the estimation methods is an important issue.Various statistical parameters can be used to measure thestrength of the statistical relationship between the esti-mated values and the reference values. I assume that vi,(i = 1, n) is the set of n reference values and ei, (i = 1, n)is the set of the estimates. �v and �e are mean of referenceand estimates values respectively. The bias, R and RMSEcan be calculated by using standard deviations of reference(rv) and estimate (re) values, mean of reference and esti-mates values, estimated values and the reference values.The bias which is the difference between the mean estimate�e and the mean reference value �v. The statistical criteria for-mula of the linear correlation coefficient R is the following,

R ¼Pn

i¼1ðvi � �vÞðei � �eÞnrvre

ð15Þ

where R measures the proximity between estimate and ref-erence. It is not sensitive to a bias (Kendall and Stuart,1963). The formula of the RMSE is;

RMSE ¼ 1

n

Xn

i¼1

ðei � viÞ2" #1

2

ð16Þ

In statistics, RMSE is a frequently used measure of thedifferences between values predicted by a model or an esti-mator and the values actually observed from the thingbeing modelled or estimated (Laurent et al., 1998).

4. Results and discussions

4.1. Estimation of land surface temperature

First, with the help of Quorum software, the data ofNOAA 12-14-15/AVHRR was converted to Level-1B for-mat which image processing software can recognize easily.Then, radiometric and geometric arrangements of theimages were done through Envi 4.3 and Idrisi Andes imageprocessing software. The primary factor that is crucial incalculating land surface temperature is brightness tempera-ture values. The brightness temperatures of the images for4th and 5th channels were obtained also through Envi 4.3and Idrisi Andes image processing software. Another fac-tor that is necessary in calculating land surface temperatureis NDVI values. To acquire the stated values; 1(VIS). and2(NIR). channels of NOAA 12-14-15/AVHRR and Eq.(6) were used, and their NDVI images were obtained.ln(NDVI) function should be used to calculate the valuesof e4 and e5 in Eq. (4) and Eq. (5). The values in the rangeof �1 and 0 will lead mathematically to trouble for

Altitude

Latitude

Longtitude

Months

Montly mean LST temperature

Monthly meanair temperature

City

Input layer Hidden layer Output layer

Fig. 4. The ANN model used in the study.

978 M. S�ahin / Advances in Space Research 50 (2012) 973–985

Page 8: Modelling of air temperature using ann and remote sensing

Author's personal copy

ln(NDVI) function. To overcome the trouble, the statedranges were deactivated in the images. The last form ofNDVI images were employed in Eq. (4) and (5), hence,emissivity values of 4th and 5th channels were acquired.By adding pre-calculated e4 and e5 values to the formulasof De = (e4 � e5) and e = (e4 + e5)/2, respectively the emis-sivity difference and mean of emissivity of 4th and 5thchannels were obtained.

In Eq. (1) land surface temperature maps were acquiredbased on Ulivieri et al. (1994) algorithm via the use ofbrightness temperature of 4th (T4) and 5th (T5) channels,De, e. In Fig. 5, land surface temperature map was shownas based on Ulivieri et al. (1994) algorithm at 06.56 a.mlocal time on 10th June 2002.

Having an overall examination over the map, it is seenthat land surface temperature on coastal lines of BlackSea Region in north of Turkey, is in the range of293–302 K whereas in internal parts of Black Sea Regionand in Eastern Regions it is within the range of

281�299 K. The temperature in Eastern Region varies,even if just a bit, between 299 and 308 K. Furthermore, itis taken in from the map that the coastal lines of Aegeanand Mediterranean Regions, the majority of Central Ana-tolia Region and inlands of Mediterranean Region taketemperature range of 296–305 K. Still, it can overtly beviewed in the map that the temperature range in South-eastern Region is mostly amongst 308–319 K. The temper-ature range in most locations of the Thrace region i.e.North-western is between 296 K and 299 K. And also,the change of temperature is seen between 278 K and287 K in cloud-covered areas. Furthermore, it is under-stood from the map that land surface temperatures ofSouth and West regions of Turkey are more than North’sand East’s regions.

At certain parts of Turkey’s neighbouring countries Iraqand Syria, the temperature map has taken the black colour.Because the stated countries have quite a weak vegetationcover and desert climate, the algorithm developed on the

Fig. 5. Land surface temperature map based on Ulivieri et al. (1994) algorithm at 06.56 a.m local time on 10th June 2002 (K).

y = 0.9991x R = 0.979

RMSE=1.778KN=3055

240

250

260

270

280

290

300

310

320

330

255 265 275 285 295 305 315

Meteorological value(K)

Sat

ellit

e va

lue(

K)

Fig. 6. The comparison of LST values obtained through Ulivieri et al. (1994) algorithm with meteorological values.

M. S�ahin / Advances in Space Research 50 (2012) 973–985 979

Page 9: Modelling of air temperature using ann and remote sensing

Author's personal copy

basis of the emissivity of vegetation was not able to havehad measurements with sufficient precision. But, becausethere is enough vegetation cover in the study areas, thetemperature measure developed based on vegetation indexalgorithm, does not bear any hinder for the study.

Similarly, the total 147 LST satellite images were derivedproviding at least one image in each month with the samemethod on the basis of algorithm developed by Ulivieriet al.(1994), between 1995 and 2005 years. Land surfacetemperature values were obtained from over the derivedimages by using coordination of 20 cities specified in Table1. By using Eqs. (15), (16), 3055 land surface temperaturevalues obtained through satellite, were compared with theones of TSMS values. At the end of the comparison, R

and RMSE were found to be 0.979 and 1.778 K, respec-tively (see Fig. 6).

Land surface temperature in various points of the worldwas calculated via satellite data. On examining the studiesin the literature, it was seen that rms error range in all stud-ies occurred between 1 K and 3 K (Price, 1984; Becker andLi, 1990; Vidal, 1991; Sobrino et al., 1996; Coll et al., 1994;Ouaidrari et al., 2002; Pinheiro et al., 2006; Katsiabaniet al., 2009; S�ahin and Kandırmaz, 2010). Because theRMSE value in the study was found as 1.778 K, the studyis in tune with the literature. So, there is no inconvenienceto use the algorithm developed by Ulivieri et al. (1994) toacquire LST values based on NOAA/AVHRR data in Tur-key. More, it is suggested researchers to use the stated algo-rithm in theirs studies.

4.2. Estimation of air temperature

In this study, ANN was employed to calculate monthlymean air temperature. The network used in the study iscomposed of input layer, hidden layer and one outputlayer. While month, altitude, latitude, longitude, city andmonthly mean land surface temperature were used asinput, monthly mean air temperature was acquired from

the output layer. Whereas the data from the period of1995–2004 were used for the training of the network, thedata of 2005 were used to test the accuracy of the trainednetwork.

There is not a mathematical formula to determine thenumber of neuron in hidden layer of ANN. The numberof neuron in hidden layer is decided in the result of net-work training. Neurons between 1 and 50 in the hiddenlayer were tested to determine the optimum artificial net-work model employed in the study. In the meantime,because the starting weights of ANNs were composedrandomly, the appropriate ANN model was decided afterthe trails.

As consequence of the trails, the values of transfer func-tions, correlation coefficients and root mean squared errorof models which were used in hidden and output layers ofthe most accurate fourteen networks were calculated (seeTable 2).

High correlation coefficient and small RMSE value is astatistical rule in the model which was developed to use incomparison of coefficient values of calculated correlationcoefficient and error mean square. After having examinedTable 2 and when air temperature estimation results werecompared to statistical criteria, it was seen that the opti-mum ANNs are 6-14-1 and 6-24-1 models which are calledM8 and M9. The transfer function in the hidden layer ofthe model 6-14-1 is tansig whereas linear in output layer.In addition, 6 neurons exist in the input layer while 14 inthe hidden layer, and 1 in the output layer. Similarly, thetransfer function in the hidden layer of the model 6-24-1,called as M9, is tansig whereas it is linear in the outputlayer. The 6 neurons exist in the input layer of the ModelM9 while 24 in the hidden layer, and 1 in the output layer.Correlation coefficients of M8 and M9 models were foundequal to one another; 0.991. That means input variables inANN got a success up to 99.1% in estimating monthlymean air temperature. But the values of RMSE are differ-ent from one another. Whereas the RMSE value of M8 is

Table 2The R and RMSE statistics of different ANN models.

Name Model Transferfunctionhidden

Transferfunctionoutput

R RMSE (K)

M1 6-05-1 Logsig Logsig 0.990 1.368M2 6-25-1 Logsig Logsig 0.989 1.471M3 6-25-1 Logsig Linear 0.987 1.562M4 6-44-1 Logsig Linear 0.968 2.403M5 6-24-1 Logsig Tansig 0.987 1.562M6 6-44-1 Logsig Tansig 0.988 1.517M7 6-50-1 Logsig Tansig 0.989 1.433M8 6-14-1 Tansig Linear 0.991 1.254M9 6-24-1 Tansig Linear 0.991 1.263M10 6-44-1 Tansig Linear 0.990 1.373M11 6-14-1 Tansig Tansig 0.989 1.430M12 6-05-1 Tansig Tansig 0.991 1.268M13 6-16-1 Tansig Logsig 0.989 1.399M14 6-34-1 Tansig Logsig 0.976 2.118

980 M. S�ahin / Advances in Space Research 50 (2012) 973–985

Page 10: Modelling of air temperature using ann and remote sensing

Author's personal copy

1.254 K, it is 1.263 K with M9. In that case, the model M8is the most accurate one developed in the study. More, byusing the model M8, city based correlation coefficient andRMSE values were calculated (see Table 3). Along withhaving rather high values on city based correlation coeffi-cient in the range of 0.966–0.997, the highest value belongsto Adana. On the other hand, the smallest value is of Kay-seri. Its RMSE values are between 0.705 K and 2.600 K. Inestimation study of monthly mean air temperature, with0.705 K, the smallest error was of Afyonkarahisar whilethe highest is of Kayseri; 2.600 K. When the study is com-pared to the other studies in the literature, it is seen that theerror ranges of the studies in the literature, for monthlymean air temperature, fluctuate between 1.79 K and 3 K(Kawamura and Edamatsu, 1993; Prihodko and Goward,1997; Czajkowski et al., 1997; Cresswell et al., 1999;Lakshmi et al., 2001; Jang et al., 2004; Riddering andQueen, 2006; Stisen et al., 2007; Vancutsem et al., 2010)whereas it is 0.705 K and 2.600 K in this study. The studyis accordant with the literature. Even, the general outcomesof the study is more accurate than the ones in the literature(R = 0.991; RMSE = 1.254 K).

Moreover in the study, estimated monthly mean air tem-perature values and city based graphical figures of monthlymeteorological values were formed (see Fig. 7).

When Fig. 7 examined, it can be seen that ANN andmeteorological values which were estimated monthly inAdana are rather close to one another. A differenceoccurred between meteorological and ANN values inSeptember and November, but through the calculationsmade, it was understood that the difference was not asmuch as thought. Whereas the error between both values

in September was 1.629 K, it was 1.002 K in November.In other months, the errors were in the range of0.020–0.671 K. When the figure examined, it will be seenthat the errors in Ankara were quite high in Februaryand December. In consequence of a statistical study, itwas emerged that an error of 1.418 K was made in thestudy carried in January for Ankara. The error reachedup to 5.825 K in February while 4.020 K in December.Similarly, the errors got values between 0.005–0.995 K inother months. It is understood from the graphic on the fig-ure that the error in August in Balıkesir was higher thanother months. As a result of calculations, it was seen thaterrors of 1.264 K in May and 2.338 K in August weremade. In other months the error range got a value between0.180–0.871 K. When the figure examined for _Izmir, it canbe seen that the amount of error in August is more thanother months. In consequence of a statistical study, itwas emerged that the error in August was found as2.903 K while in other months between 0.106 and0.965 K. As can be seen from the figure, the amount oferror in Samsun was high in February, October andDecember; 1.514 K in February, 1.729 K in October and2.231 K in December. In other months, the error valueswere in the range of 0.013–0.913 K. It was seen that theerror value for S�anlıurfa was rather high in June whereaslow in October and December. The error was found as3.990 K in June while 1.216 K in October, and 1.012 K inDecember. The error range acquired in other months wasbetween 0.022 K and 0.848 K. It can be understood fromthe figure that the amount of error in Van was rather highin January, February, November, and December. Theerror ranges were 3.303 K in January; 2.275 K in February;1.533 K in November; and 1.122 K in December whilebetween 0.226 and 0.665 K in other months.

Generally describing, it is understood that the cities ofAfyonkarahisar, Antalya, Artvin, Erzurum, Eskis�ehir,Kars, and Sivas have rather close meteorological andmonthly air temperature values with one another. Errorranges of the cities resulted in below 1.000 K. It can beunderstood from the figure that error values of Denizliincreased in May and September. The error value found2.100 K in May while 1.400 K in September. The errorvalue in _Istanbul-Goztepe found between 1.200–1.800 Kin May, June, September, October, November, andDecember whereas the value stayed under 1.000 K in othermonths. In Kayseri, although meteorological and esti-mated values happened fairly close, the error values werein the range of 1.200–8.800 K in the months of June andJuly, respectively. The error value of Konya fluctuatedbetween 1.300 and 3.500K in January, February, July,August, and November whereas the value stayed under1.000 K in the other months. Although the error valuesgot values below 1 K in Malatya in March, October,November, and December, the error values in othermonths were between 1.000 and 2.400 K. The error valueof 0.200–0.900 K prevailed in March, April, November,

Table 3The correlation coefficient and RMSE values of cities for monthly meanair temperature.

Stations Correlationcoefficient (R)

RMSE (K)

Adana 0.997 0.767Afyonkarahisar 0.995 0.705Ankara 0.983 2.142Antalya 0.993 0.904Artvin 0.995 0.779Balıkesir 0.994 0.991Denizli 0.990 1.060Erzurum 0.996 0.961Eskis�ehir 0.995 0.826_Istanbul-Goztepe 0.991 1.134Izmir 0.992 0.993Kars 0.996 0.733Kayseri 0.966 2.600Konya 0.992 1.520Malatya 0.995 1.477Rize 0.990 1.189Samsun 0.990 0.987Sivas 0.994 0.880S�anlıurfa 0.993 1.309Van 0.991 1.343

M. S�ahin / Advances in Space Research 50 (2012) 973–985 981

Page 11: Modelling of air temperature using ann and remote sensing

Author's personal copy

and December in the city of Rize while the values changedfrom 1.000 to 2.100 K in the other months.

5. Conclusion

In the study, ANN and remote sensing methods wereused to estimate monthly mean air temperature in Adana,

Afyonkarahisar, Ankara, Antalya, Artvin, Balıkesir, Den-izli, Erzurum, Eskis�ehir, _Istanbul-Goztepe, _Izmir, Kars,Kayseri, Konya, Malatya, Rize, Samsun, Sivas, S�anlıurfaand Van. While month, latitude, longitude, altitude, cityand monthly mean land surface temperature obtainedthrough satellite data were used as input in ANN, monthlymean air temperature were used as output. Satellite based

Adana

270

280

290

300

310

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Meteorological value(K)

Estimated value(K)

Afyonkarahisar

260

270

280

290

300

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Ankara

250

260

270

280

290

300

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Antalya

270

280

290

300

310

1 2 3 4 5 6 7 8 9 10 11 12

MonthsA

ir te

mpe

ratu

re(K

)

Artvin

260

270

280

290

300

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Balıkesir

260

270

280

290

300

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Denizli

270

280

290

300

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Erzurum

240

250

260

270

280

290

300

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Eskişehir

260

270

280

290

300

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

İzmir

270

280

290

300

310

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Fig. 7. Comparison of monthly mean air temperature between ANN and meteorological values.

982 M. S�ahin / Advances in Space Research 50 (2012) 973–985

Page 12: Modelling of air temperature using ann and remote sensing

Author's personal copy

land surface temperature values were obtained through theuse of NOAA-AVHRR satellite data together with Ulivieriet al. (1994) algorithm. The values were compared staticallyto the meteorological values. In accordance with the litera-ture, R and RMSE values were found 0.979 and 1.778 K,respectively.

Various ANN models were employed in monthly meanair temperature estimation. The data from the period of1995–2004 were used to develop the models while the2005 data were used to test the trained models. The opti-mum result was obtained through the model 6-14-1. Sixneurons exist in the input layer of this model while four-

İstanbul-Göztepe

260

270

280

290

300

310

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Kars

250

260

270

280

290

300

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Kayseri

260

270

280

290

300

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Konya

260

270

280

290

300

1 2 3 4 5 6 7 8 9 10 11 12

MonthsA

ir te

mpe

ratu

re(K

)

Malatya

260

270

280

290

300

310

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Rize

260

270

280

290

300

310

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Samsun

260

270

280

290

300

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Sivas

260

270

280

290

300

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Şanlıurfa

260

270

280

290

300

310

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Van

260

270

280

290

300

1 2 3 4 5 6 7 8 9 10 11 12

Months

Air

tem

pera

ture

(K)

Fig 7. (continued)

M. S�ahin / Advances in Space Research 50 (2012) 973–985 983

Page 13: Modelling of air temperature using ann and remote sensing

Author's personal copy

teen in the hidden layer, and one in the output layer.More, R and RMSE values were found as 0.991 and1.254 K, respectively when estimation results of themodel were compared to meteorological values. The out-comes obtained through the study shows that the presentstudy is more accurate than other studies in theliterature.

To develop a successful model is very important like instudy belongs to remote sensing and ANN. Because, even ifit is possible for the same study to be successful only basedon meteorological data, inadequacy of meteorological sta-tions, financial burden, and inconvenient geographic distri-bution of meteorological stations due to climate conditionscause remote sensing method to be unavoidable. Asknown, meteorological satellites employed in remote sens-ing studies are able to commit easier and cheaper datatransmitting by scanning the land surface. Thus, ANNtechnique and remote sensing can be used as an alternatemethod in air temperature estimations which are usefulfor lots of applications including study of vector-borne dis-eases, epidemic forecasting, weather forecasting, veterinaryuses, climate change, determination of various heat andradiation fluxes, vapour pressure deficit, water potential,urban land use and urban heat island, shortwave and long-wave radiation, stomatal resistance, ecology, hydrologyand atmospheric sciences. And also, this method is offeredto researchers who study for the health of human being,especially about high blood pressure, ischemic heart dis-ease, respiratory infections and system, heat related mor-tality, influenza, ambient temperature.

References

Aasamaa, K., Sober, A. Stomatal sensitivities to changes in leaf waterpotential, air humidity, CO2 concentration and light intensity and theeffect of abscisic acid on the sensitivities in six temperate deciduous treespecies. Environ. Exp. Bot. 71, 72–78, 2011.

Almeida, S.P., Casimiro, E., Calheiros, J. Effects of apparent temperatureon daily mortality in Lisbon and Oporto. Environ. Health 9, 1–7,Portugal, 2010.

Analitis, A., Katsouyanni, K., Biggeri, A., Baccini, M., Forsberg, B.,Bisanti, L., Kirchmayer, U., Ballester, F., Cadum, E., Goodman, P.G.,Hojs, A., Sunyer, J., Tiittanen, P., Michelozzi, P. Effects of coldweather on mortality: results from 15 european cities within thePHEWE project. Am. J. Epidemiol. 168, 1397–1408, 2008.

Ballester, F., Corella, D., Perez-Hoyos, S., Saez, M., Hervas, A. Mortalityas a function of temperature: a study in Valencia, Spain, 1991–1993.Int. J. Epidemiol. 26, 551–561, 1997.

Becker, F. The impact of spectral emissivity on the measurement of landsurface temperature from a satellite. Int. J. Remote Sens. 8, 1509–1522,1987.

Becker, F., Li, Z.-L. Surface temperature and emissivity at various scales:definition, measurement and related problems. Remote Sens. Rev. 12,225–253, 1995.

Becker, F., Li, Z.L. Toward a local split window method over landsurface. Int. J. Remote Sens. 11, 369–393, 1990.

Bian, L., Li, L., Yan, G. Combining global and local estimates for spatialdistribution of mosquito larval habitats. GISci. Remote Sens. 43, 128–141, 2006.

Bocchiola, D., Diolaiuti, G. Evidence of climate change within theAdamello Glacier of Italy. Theor. Appl. Climatol. 100, 351–369, 2010.

Brunsell, N.A., Mechem, D.B., Anderson, M.C. Surface heterogeneityimpacts on boundary layer dynamics via energy balance partitioning.Atmos. Chem. Phys. 11, 3403–3416, 2011.

Cam, E., Arcaklıoglu, E., Cavus�oglu, A., Akbıyık, B.A. Classificationmechanism for determining average wind speed and power in severalregions of Turkey using artificial neural networks. Renew. Energy 30,227–239, 2005.

Cheval, S., Dumitrescu, A., Bell, A. The urban heat island of Bucharestduring the extreme high temperatures of July 2007. Theor. Appl.Climatol. 97, 391–401, 2009.

Coll, C., Sobrino, J.A., Valor, E. On the atmospheric dependence of thesplit-window equation for land surface temperature. Int. J. RemoteSens. 15, 105–122, 1994.

Cresswell, M.P., Morse, A.P., Thomson, M.C., Connor, S.J. Estimatingsurface air temperatures from METEOSAT land surface temperaturesusing an empirical solar zenith angle model. Int. J. of Remote Sens. 20,1125–1132, 1999.

Czajkowski, K.P., Mulhern, T., Goward, S.N., Cihlar, J., Dubayah, R.O.,Prince, S.D. Biospheric environmental monitoring at BOREAS withAVHRR observations. J. Geophys. Res. 102, 29651–29662, 1997.

Elwood, P.C., Beswick, A., O’Brien, J.R., Renaud, S., Fifield, R., Limb,E.S., Bainton, D. Temperature and risk factors for ischaemic heartdisease in the caerphilly prospective study. British Heart J. 70, 520–523, 1993.

Gates, D.M. Biophysical Ecology. Springer, New York, 1980.Goetz, S.J., Prince, S.D., Small, J. Advances in satellite remote sensing of

environmental variables for epidemiological applications. Adv. Para-sit. 47, 289–307, 2000.

Goward, S.N., Markham, B., Dye, D.G., Dulaney, W., Yang, J.Normalized difference vegetation index measurements from theadvanced very high resolution radiometer. Remote Sens. Environ.35, 257–277, 1991.

Haykin, S. Neural Networks: A Comprehensive Foundation. Prentice-Hall International Inc., New Jersey, 1999.

Heding, S., Kai, L., Hanchun, C., Xianlong, C., Yongan, H., Zhiyi, S.Experimental ecology and hibernation of onchidium struma (gastrop-oda: pulmonata: systellommatophora). J. Exp. Mar. Biol. Ecol. 396,71–76, 2011.

Jain, S.K., Jain, S.K., Hariprasad, V., Choudhry, A. Water balance studyfor a basin integrating remote sensing data and GIS. Indian Society ofRemote Sensing, 2011, http://dx.doi.org/10.1007/s12524-011-0078-2

Jang, J.D., Viau, A.A., Anctil, F. Neural network estimation of airtemperatures from AVHRR data. Int. J. Remote Sens. 25, 4541–4554,2004.

Kalogirou, S.A. Applications of artificial neural-networks for energysystems. Appl. Energy 67, 17–35, 2000.

Kalogirou, S.A. Artificial neural networks in renewable energy systemsapplications: a review. Renew. Sust. Energy Rev. 5, 373–401, 2001.

Katsiabani, K., Adaktilou, N., Cartalis, C. A generalised methodology forestimating land surface temperature for non-urban areas of Greecethrough the combined use of NOAA–AVHRR data and ancillaryinformation. Adv. Space Res. 43, 930–940, 2009.

Kawamura, H., Edamatsu, Y. Better understanding of earth environment.Int. Geosci. Remote Sens. Symp. 3, 473–1475, 1993.

Kendall, M.A., Stuart, A. The Advanced Theory of Statistics. Griffin,London, 1963.

Kittel, T.G.F., Baker, B.B., Higgins, J.V., Haney, J.C. Climate vulnera-bility of ecosystems and landscapes on Alaska’s North Slope. Reg.Environ. Change 11, 249–264, 2011.

Kucharik, C.J., Serbin, S.P., Vavrus, S., Hopkins, E.J., Motew, M.M.Patterns of climate change across Wisconsin from 1950 to 2006. Phys.Geogr. 31, 1–28, 2010.

Kunst, A.E., Looman, C.W.N., Johan, P.M. Outdoor air temperature andmortality in the Netherlands: a time-series analysis. Am. J. Epidemiol.137, 331–334, 1993.

Lakshmi, V., Czajkowski, K., Dubayah, R., Susskind, J. Land surface airtemperature mapping using TOVS and AVHRR. Int. J. Remote Sens.22, 643–662, 2001.

984 M. S�ahin / Advances in Space Research 50 (2012) 973–985

Page 14: Modelling of air temperature using ann and remote sensing

Author's personal copy

Laurent, H., Jobard, I., Toma, A. Validation of satellite and ground-basedestimates of precipitation over the Sahel. Atmos. Res. 47–48, 651–670,1998.

Lee, S.H., Kim, S.W., Angevine, W.M., Bianco, L., McKeen, S.A., Senff,C.J., Trainer, M., Tucker, S.C., Zamora, R.J. Evaluation of urbansurface parameterizations in the WRF model using measurementsduring the Texas air quality study 2006 field campaign. Atmos. Chem.Phys. 11, 2127–2143, 2011.

Michelozzi, P., Accetta, G., De Sario, M., D’Ippoliti, D., Marino, C.,Baccini, M., Biggeri, A., Anderson, H.R., Katsouyanni, K., Ballester,F., Bisanti, L., Cadum, E., Forsberg, B., Forastiere, F., Goodman,P.G., Hojs, A., Kirchmayer, U., Medina, S., Paldy, A., Schindler, C.,Sunyer, J., Perucci, C.A. High temperature and hospitalizations forcardiovascular and respiratory causes in 12 European cities. Am. J.Respir. Crit. Care Med. 179, 383–389, 2009.

Myint, S.W., Brazel, A., Okin, G., Buyantuyev, A. Combined effects ofimpervious surface and vegetation cover on air temperature variationsin a rapidly expanding desert city. GISci. Remote Sens. 47, 301–320,2010.

Ouaidrari, H., Gowarda, S.N., Czajkowskib, K.P., Sobrinoc, J.A.,Vermotea, E. Land surface temperature estimation from AVHRRthermal infrared measurements: An assessment for the AVHRR LandPathfinder II data set. Remote Sens. Environ. 81, 114–128, 2002.

Pinheiro, A.C.T., Mahoney, R., Privette, J.L., Tucker, C.J. Developmentof a daily long term record of NOAA-14 AVHRR land surfacetemperature over Africa. Remote Sens. Environ. 103, 153–164, 2006.

Price, J.C. Land surface temperature measurements from the split windowchannels of the NOAA-7/AVHRR. J. Geophys. Res. 89, 7231–7237,1984.

Prihodko, L., Goward, S.N. Estimation of air temperature from remotelysensed surface observations. Remote Sens. Environ. 60, 335–346, 1997.

Riddering, J.P., Queen, L.P. Estimating near-surface air temperature withNOAA –AVHRR. Can. J. Remote Sens. 32, 33–43, 2006.

Santos, P., Negri, A.J. A comparison of the normalized differencevegetation index and rainfall for the Amazon and Northeastern Brazil.J. Appl. Meteorol. 36, 958–965, 1997.

Smith, L.C. Agents of change in the New North. Eurasian Geography andEconomics. 52, 30–55, 2011.

Sobrino, J.A., Li, Z.L., Stoll, M.P., Becker, F. Multi-channel and multi-angle algorithms for estimating sea and land surface temperature withATSR data. Int. J. Remote Sens. 17, 2089–2114, 1996.

Stanelle, T., Vogel, B., Vogel, H., Baumer, D., Kottmeier, C. Feedbackbetween dust particles and atmospheric processes over West Africaduring dust episodes in March 2006 and June 2007. Atmos. Chem.Phys. 10, 10771–10788, 2010.

Stisen, S., Sandholt, I., Nørgaard, A., Fensholt, R., Eklundh, L.Estimation of diurnal air temperature using MSG SEVIRI data inWest Africa. Remote Sens. Environ. 110, 262–274, 2007.

S�ahin, M., Kandırmaz, H.M. Calculation land surface temperaturedepending on Becker and Li-1990 algorithm. J. Therm. Sci. Technol.30, 35–43, 2010.

Thomson, M.C., Connor, S.J., Milligan, P.J.W., Flasse, S. The ecology ofmalaria – as seen from earth observation satellites. Ann. Trop. Med.Parasit. 90, 243–264, 1996.

Ulivieri, C., Castronuovo, M.M., Francioni, R., Cardillo, A. A splitwindow algorithm for estimating land surface temperature fromsatellites. Adv. Space Res. 14, 59–65, 1994.

Valor, E., Casselles, V. Mapping of land surface emissivity from NDVI:application to European, African, and South American Areas. RemoteSens. Environ. 57, 167–184, 1995.

Vancutsem, C., Ceccato, P., Dinku, T., Connor, S.J. Evaluation ofMODIS land surface temperature data to estimate air temperature indifferent ecosystems over Africa. Remote Sens. Environ. 114, 449–465,2010.

Vazquez, D.P., Reyes, F.J.O., Arboledas, L.A. A comparative study ofalgorithms for estimation of land surface temperature from AVHRR.Remote Sens. Environ. 62, 215–222, 1997.

Vidal, A. Atmospheric and emissivity correction of land surface temper-ature measured from satellite using ground measurements or satellitedata. Int. J. Remote Sens. 12, 2449–2460, 1991.

Viotti, P., Liuti, G., Genova, P.D. Atmospheric urban pollution:applications of an artificial neural network (ANN) to the City Peugia.Ecol. Model. 148, 27–46, 2002.

Vogl, T.P., Mangis, J.K., Rigler, A.K., Zink, W.T., Alkon, D.L.Accelerating the convergence of the backpropagation method. Biol.Cybern. 59, 257–263, 1988.

Yu, Y., Privette, J.L., Pinheiro, A.C., Vogel, R. Theoretical evaluation ofsplit window methods to retrieve land surface temperature. IEEETrans. Geosci. Remote Sens. 46, 179–192, 2008.

M. S�ahin / Advances in Space Research 50 (2012) 973–985 985